Does Fine-Tuning Still Pay Once a Stronger Planner Starts?

Prior studies found that written plans from planners no larger than the agent shrink its fine-tuning gain; here the planner is a stronger model that taught the agent. The smaller agent, Granite 4.2 8B, tuned on gpt-5.6-luna's runs of AppWorld training tasks, outscored its off-the-shelf version. On AppWorld's harder test tasks, its lead in the share of automatic checks passed was 26.43 percentage points without help and, unlike in those studies, survived the planner's written plan and the planner acting every 10 steps. A head start, replaying the planner's first 6 actions, brought it to -1.60 [-5.07, +1.85] on runs the smaller agent took over. Planned in advance: On harder tasks, 6 actions removed three quarters; no 5-point lead remained. This also held against an off-the-shelf agent hinted to use the user's passwords and never sign up. On tasks fully completed, the head start moved the lead from +7.67 [+5.04, +10.67] to -6.36 [-10.62, -2.21]. We hypothesise that much of the lead sits in a learned opening routine, though head starts may even out any two agents.